It's untenable to simply trust LLM API providers that the models they are serving through an API endpoint is the model they claim it is. They could easily switc

It's untenable to simply trust LLM API providers that the models they are serving through an API endpoint is the model they claim it is. They could easily switch the model with a cheaper one whenever they wanted, and since LLM outputs are non-deterministic (presuming a random seed), it would be impossible to prove this. LLM's integrated into any real product requires a model hash, otherwise the provider of the model has full control over any sort of deception.